Reproduce “Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuyang You, Yongzhi Li, Jiahui Li, Yadong Mu, Quan Chen, Peng Jiang
Researchers from Peking University and Kuaishou Technology developed an adaptive video distillation framework to counter color oversaturation and temporal collapse in few-step video generation. This framework enables efficient, high-quality video synthesis, achieving top scores on VBench2 and VBench1 benchmarks while reducing inference time by 30%.
Key insights
- Adaptive weighting in a regression loss is crucial for stable distillation, allowing the student model to learn from reliable real data regions while mitigating artifacts from distribution shifts and supporting direct fine-tuning.
- Explicitly regularizing temporal variance during distillation is essential to preserve and enhance motion dynamics in generated videos, overcoming the issue of static or stiff outputs.
- Decoupling inference steps by performing initial, high-noise denoising at a reduced frame rate, followed by latent space interpolation, can substantially reduce computational load without sacrificing video quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.